Improve shape inference of flatten Op (#314)

* change code

Signed-off-by: chentong <chentong@us.ibm.com>

* add test

Signed-off-by: chentong <chentong@us.ibm.com>

* fix type error

Signed-off-by: chentong <chentong@us.ibm.com>

Co-authored-by: Alexandre Eichenberger <alexe@us.ibm.com>

src/Dialect/ONNX/ONNXOps.cpp

@@ -2153,32 +2153,48 @@ LogicalResult ONNXSplitOp::inferShapes() {
 //===----------------------------------------------------------------------===//
 
 LogicalResult ONNXFlattenOp::inferShapes() {
-  assert(axis() == 1 && "ONNXFlattenOp can only handle axis=1 for now");
   auto inTy = input().getType().dyn_cast<ShapedType>();
   if (!inTy) {
     return emitOpError("Input is a non-shaped type");
   }
-  auto outTy = output().getType().dyn_cast<ShapedType>();
+
-  if (!outTy) {
+  auto axisValue = axis();
-    return emitOpError("Output is a non-shaped type");
+  auto inputShape = inTy.getShape();
+  auto inputRank = inputShape.size();
+  if (axisValue < -1 * (int64_t)inputRank || axisValue > (int64_t)inputRank) {
+    return emitOpError("ONNXFlattenOP: axis() value is out of range");
   }
 
-  // TODO(tjingrant): Seems like we can also fairly easily support the case
+  SmallVector<int64_t, 2> dims;
-  //                  where the batch dimension is dynamic
+
-  if (!outTy.hasStaticShape()) {
+  // Negative axis is counting dimension from back
-    auto inShape = inTy.getShape();
+  if (axisValue < 0)
-    assert(inShape.size() >= 1 && "ONNXFlattenOp inShape.size() should be > 0");
+    axisValue = inputRank + axisValue + 1;
-    uint64_t outDim = 1;
+
-    for (auto it = inShape.begin() + 1; it < inShape.end(); it++) {
+  // Determine the size of the first dimension of output
-      outDim *= *it;
+  int64_t firstDim = 1;
+  for (auto i = 0; i < axisValue; i++) {
+    if (inputShape[i] == -1) {
+      firstDim = -1;
+      break;
     }
-
+    firstDim *= inputShape[i];
-    SmallVector<int64_t, 2> dims;
-    // https://pytorch.org/docs/master/generated/torch.nn.Flatten.html
-    dims.emplace_back(inShape[0]);
-    dims.emplace_back(outDim);
-    getResult().setType(RankedTensorType::get(dims, outTy.getElementType()));
   }
+  dims.emplace_back(firstDim);
+
+  // Determine the size of the second dimension of output
+  int64_t secondDim = 1;
+  for (auto i = axisValue; i < inputRank; i++) {
+    if (inputShape[i] == -1) {
+      secondDim = -1;
+      break;
+    }
+    secondDim *= inputShape[i];
+  }
+  dims.emplace_back(secondDim);
+
+  // Set the type of output
+  getResult().setType(RankedTensorType::get(dims, inTy.getElementType()));
 
   return success();
 }

test/mlir/onnx/onnx_shape_inference.mlir

@@ -639,6 +639,40 @@ func @test_flatten_1(%arg0 : tensor<5x2x3x4xf32>) -> tensor<*xf32> {
   // CHECK: return [[RES]] : tensor<5x24xf32>
 }
 
+// -----
+
+// Test when axis is 0
+func @test_flatten_2(%arg0 : tensor<2x3x4xf32>) -> tensor<*xf32> {
+  %1 = "onnx.Flatten"(%arg0) {axis = 0 : si64} : (tensor<2x3x4xf32>) -> tensor<*xf32>
+  "std.return"(%1) : (tensor<*xf32>) -> ()
+  // CHECK-LABEL: test_flatten_2
+  // CHECK: [[RES:%.+]] = "onnx.Flatten"(%arg0) {axis = 0 : si64} : (tensor<2x3x4xf32>) -> tensor<1x24xf32>
+  // CHECK: return [[RES]] : tensor<1x24xf32>
+}
+
+// -----
+
+// Test when axis is negative
+func @test_flatten_3(%arg0 : tensor<2x3x4xf32>) -> tensor<*xf32> {
+  %1 = "onnx.Flatten"(%arg0) {axis = -1 : si64} : (tensor<2x3x4xf32>) -> tensor<*xf32>
+  "std.return"(%1) : (tensor<*xf32>) -> ()
+  // CHECK-LABEL: test_flatten_3
+  // CHECK: [[RES:%.+]] = "onnx.Flatten"(%arg0) {axis = -1 : si64} : (tensor<2x3x4xf32>) -> tensor<24x1xf32>
+  // CHECK: return [[RES]] : tensor<24x1xf32>
+}
+
+// -----
+
+// Test when input is not static shape
+func @test_flatten_4(%arg0 : tensor<2x4x5x?xf32>) -> tensor<*xf32> {
+  %1 = "onnx.Flatten"(%arg0) {axis = 2 : si64} : (tensor<2x4x5x?xf32>) -> tensor<*xf32>
+  "std.return"(%1) : (tensor<*xf32>) -> ()
+  // CHECK-LABEL: test_flatten_4
+  // CHECK: [[RES:%.+]] = "onnx.Flatten"(%arg0) {axis = 2 : si64} : (tensor<2x4x5x?xf32>) -> tensor<8x?xf32>
+  // CHECK: return [[RES]] : tensor<8x?xf32>
+}
+
+
 //===----------------------------------------------------------------------===//
 /// Test the reshape op inference when concat are present.
 //===----------------------------------------------------------------------===//